Multiple Robust Controller Design based on Parameter Dependent Lyapunov Functions

Multiple Robust Controller Design based on Parameter Dependent Lyapunov Functions Jongeun Choi, Ryozo Nagamune and Roberto Horowitz Abstract This paper tackles the problem of simultaneously designing a partition of an uncertain set and its corresponding set of multiple controllers that optimize the worst-case performance of a linear time invariant system under parametric uncertainty. The parametric uncertainty region is assumed to be convex polytopic, which is also partitioned into a set of convex polytopic local regions. It is desired that all plants that belong to a local region are to be controlled by a single controller, which is designed, based on a parameter dependent Lyapunov function, to give an optimal worst-case performance for that region. The total performance is evaluated as the maximum of worst-case performances for all the local regions. The performance is minimized with respect to a fixed number of convex polytopic local regions, as well as the same number of controllers. Even though the formulated problem is nonconvex, and thus it is difficult to ensure global optimality, descent algorithms are provided to update the local regions and the multiple controllers so that they guarantee monotonic non-increasing total performance. I. INTRODUCTION Controller robust performance is of great importance when there is uncertainty in a plant to which performance is sensitive. It is normally quantified as the worst-case performance that is attained by the control system under a modeled plant uncertainty. Many controller analysis and synthesis methodologies have been obtained to achieve both robust H 2 and robust H performance with a single fixed control system, mainly based on numerically efficient convex optimization (or LMI) techniques. Most results on robust performance synthesis so far have been concerned with the design of a single controller for all plants in the assumed model set. However, a single controller may not be sufficient to achieve the required performance, since the robustness requirement has to be met at the expense of performance limitations. Moreover, many standard controller design techniques, which find both a single Lyapunov function and a single controller, may be infeasible if the size of uncertainty is large. To overcome these drawbacks, in this paper, we will consider the design of a set of controllers that guarantee robust performance for parametric uncertain systems. Each controller in this set takes charge of a local region in the Jongeun Choi and Roberto Horowitz are with the Department of Mechanical Engineering, University of California at Berkeley {jchoi, horowitz}@me.berkeley.edu Ryozo Nagamune is with the Department of Mathematics, the Royal Institute of Technology in Sweden. ryozo@kth.se entire uncertainty region and is designed to give an optimal worst-cast performance for the specific region for which it is designed. The total performance of the control set is evaluated as the maximum of such worst-case performance in each local region. Therefore, the total performance of this set of multiple controllers is expected to be better than the worst-case performance of a single controller, designed for the entire region. In addition, segmenting an uncertainty region into smaller regions and utilizing multiple robust controllers is advantageous from the controller design perspective, since it is generally easier to obtain a global Lyapunov function for an uncertainty region, as the region size decreases. Systems with multiple controllers have increasingly drawn the attention of control engineers in adaptive control or multiple modeling applications, in order to improve the controller performance under very different conditions 2, 3, 4. However, most of the results so far have not paid enough attention to the importance on how to divide the entire uncertainty (or a set of different situations) into small partitioned regions. Here, we study how to simultaneously determine a partition of the uncertainty and design its corresponding a set of robust controllers, in order to improve the total performance of the control system. In this paper, minimization of the total performance is considered with respect to the fixed number of partitioned polytopic uncertainty subsets and controllers. Each partitioned uncertainty region is enforced to be convex polytopic, and the controller of each partition is designed based on a parameter dependent Lyapunov function and a convex optimization technique. Unfortunately, the minimization problem turns out to be hard. Even for a given set of convex polytopic uncertainty regions, designing globally optimal full-order output-feedback controllers for polytopic uncertain systems is a nonconvex optimization problem 5. Therefore, we present algorithms for updating both a partition of the uncertainty set and the set of multiple controllers that guarantee monotonic non-increasing total performance. The resulting partition of an uncertainty set and robust controllers can be utilized in the following way. For each real plant, uncertain parameters are to be estimated either off-line or on-line. For off-line estimation, a local region that the estimated parameters lie will be detected, and a controller corresponding to that region will be selected and applied to the real plant. On the other hand, for on-

line estimation, we need to use an initial robust controller designed for the entire uncertainty region during parameter identification 6, 7. After finding the local region that contains an estimated parameter, a user can switch the controller to the particular controller that was specially designed for that region, to improve performance. This paper is organized as follows. In Section II, we introduce several concepts and notations concerning uncertain systems that are necessary for the subsequent arguments. Using these concepts, Section III formulates a worst-case performance minimization problem with multiple controllers. For the formulated problem, in Section IV, descent algorithms are presented to design a partition and its multiple robust controllers simultaneously, which are sub-optimal. The algorithms intertwine partitioning techniques of polytopic uncertain systems and a robust controller design method, both of which are also explained in Section IV. In Section V, we apply the algorithm to an example in order to illustrate the potential of multiple controllers with a properly chosen partition compared to a single controller. The initial robust H controller synthesis, which is important to our nonconvex optimization, is described in Appendix. II. PRELIMINARIES In this section, we define and review several concepts that will arise throughout the paper. A. Notations R n m is the set of real n m matrices. R n is the set of real n dimensional vectors. R + is the set of positive real numbers. Z + is the set of positive integers. The interior of a set S is denoted as Int(S). The Cartesian product of the indexed family of sets {S i } is denoted by i Z + S i. The symbol Sym( ) denotes Sym(A) = A + A, where A is a complex conjugate transpose of A. Co {A,, A m } represents the convex hull of a set of matrices {A i } and is defined as the intersection of all convex sets containing all elements of {A i }. V(S) is the set of all the vertices of a convex polytope S. The direct sum of two matrices A and B is denoted as A B := diag(a, B). The Kronecker product of two matrices A R n p with entries a ij and B R m k is denoted by A B and defined by the mn kp block matrix A B := a B a p B..... a n B a np B. () A realization of the transfer function of a discrete time LTI system is denoted by A B := D + C(zI A) B. (2) C D z y Generalized Plant K Fig.. A generalized plant with parametric uncertainty, (denoted by Σ ) and a controller K. B. A Partition of a Parameter Set For a given compact and osed parameter set Λ, we call {Λ q : q Z N } a partition of a set Λ, where Z N is the set of N coordinates of the partition, if Λ q Λ is nonempty, osed and satisfies Σ Int(Λ i ) Int(Λ j ) = for i j, Λ q = Λ. (3) In words, a partition of a set Λ is a collection of nonempty subsets of Λ whose interior is disjoint and whose union is all of Λ. C. Probability of a Stochastic Process Define a random vector from a stochastic process : Z + Λ with a probability density function f : Λ R +. The probability of being contained in ˆΛ Λ is denoted as Prob( ˆΛ) and defined by the formula Prob( ˆΛ) := f (µ)dµ. (4) µ ˆΛ D. Polytopic Parametric Uncertain LTI Systems Let us consider a discrete time linear time invariant generalized plant with parametric uncertainty Λ, in short denoted by Σ, as depicted in Fig., z y = A B B 2 C D D 2 C 2 D 2 w u w u, A R n n, where z R nz is the output of the system, w R nw is the disturbance to the system, u R nu is the control action and y R ny is the measured output. Here, the superscript on a matrix means that the matrix is a function of an uncertain parameter vector. The set of all uncertain plants with Λ is denoted as Σ Λ given by the formula (5) Σ Λ := { Σ : Λ }. (6) We only consider the discrete time case here but the application to the continuous time case is straightforward.

Assumptions on the plant parameters are as follow. A. (A, B 2, C 2 ) is stabilizable and detectable for each Λ, A.2 Λ is a convex polytope. The first assumption A. is necessary and sufficient for each system in Σ Λ to be stabilizable with dynamic output feedback. A.2 is necessary for robust controller design based on LMI techniques. E. Convexity of Parametric Uncertain LTI Systems Consider a set of LTI systems { } A B : Λ C D where Λ R u. Correspondingly, we define the set of system matrices as { G : Λ } { } A B := : Λ. (8) C D We assume that R u and G R n n2, where n := n + n z + n y and n 2 := n + n w + n u is given by G = G + ( T I n ) G. G u (7). (9) Affine parameter-dependent models 8 are readily converted to polytopic ones. If the matrix G is affine in, it maps a convex polytopic set Λ to a set of convex polytopic system matrices. This is important for us to synthesize a robust controller based on LMI optimization techniques. The following lemma is used for obtaining polytopic uncertain systems. Lemma : If G is affine in, and if Λ is a convex polytope, then the set of all the matrices { G : Λ } in Eq. (9) is a convex polytope in R n n2, and equivalent to the following set Φ Λ Φ Λ := Co { G : V(Λ) }. () The proof is straightforward and thus omitted here. III. PROBLEM FORMULATION Consider the set Σ Λ of parametric uncertain LTI systems, each of which is represented in Eq. (5). We would like to simultaneously determine ) a partition {Λ q : q Z N } of Λ, and 2) synthesize a set of robust controllers {K q : q Z N }, where K q is applied to the set Σ Λq, such that a certain worst-case performance cost (to be defined below) is minimized. The set of robust controller s coordinates is denoted as Z N and let the coordinate of a robust controller be also the coordinate of the corresponding subset of the partition. The regional controller K q will be designed to regulate any linear uncertain system in the uncertain plant set Σ Λq. Given a proper real-rational controller K q (z) for uncertain systems Σ Λq AKq B K q (z) := Kq, A Kq R n n, () C Kq D Kq a realization of the osed loop transfer function from disturbance signal w to the output signal z is denoted by T zw (, K q ). We define several performance measures. Definition 2: (Performance cost) For a given set of uncertain LTI systems Σ Λq and a robustly stabilizing controller K q, we define the worst case performance cost of the uncertain osed-loop systems by where J i (Λ q, K q ) := sup Λ q T zw (, K q ) i, { i = 2, H2 norm performance, i =, H norm performance. (2) In the case when the probability density function f is given, then the averaged stochastic worst-case performance cost can be defined by EJ i (Λ q, K q ) := sup T zw (, K q ) i f ()d, Λ q Λ q (3) The total performance is defined as the maximum of such worst-case performance of each local region given by max J i (Λ q, K q ), (4) for the worst-case performance case and by max EJ i (Λ q, K q ), (5) for the averaged stochastic worst-case performance case. Remark 3: From now on, we will deal mainly with the H performance (the induced L 2 system gain) case, and the worst-case performance cost in Eq. (2). However, the results for the H 2 performance case (or even mixed H 2 /H performance case 9) and/or the averaged stochastic worst-case performance cost in Eq. (3) can be dealt with in an analogous manner. We are interested in minimizing the total performance, which sets an upper bound on the performance of an individual osed-loop system. Next, the total performance minimization problem is formally stated. Problem: For a given set of uncertain LTI systems Σ Λ with a convex polytope Λ, design a partition {Λ q : q Z N } of the uncertainty set Λ and a set of stabilizing robust controllers {K q : q Z N }, such that the worst-case H performance cost J (Λ q, K q ) in Eq. (2) for Λ is minimized, i.e., solve the optimization problem min {Λ q},{k q K(Λ q)} max J (Λ q, K q ), (6) where K(Λ q ) is the set of all controllers that internally stabilize the osed-loop system for all Λ q.

The formulated optimization is nonconvex and therefore, it is difficult to solve exactly in the global optimal sense. In the following section, we will explain how to get a reasonable solution in a systematic way. IV. DESIGN OF A PARTITION AND MULTIPLE CONTROLLERS First we show how to partition a convex polytopic uncertainty set into a set of smaller convex polytopic uncertainty regions using the Cartesian product. The imposition of the constraint on a partition that it must have convex polytopic subsets, simplifies the tasks of determining partitions of uncertain systems as well as their respective sets of multiple robust controllers. Subsequently, we discuss the design of a set of robust controllers. Finally, a descent search algorithm for the optimization given by Eq. (6) will be presented. The sampled uncertain system can be interpreted as follows. Recall : Z + Λ is an uncertain parameter of a linear system. For example, we may think of i Z + as a serial number of the product (or plant) in the manufacturing line. We assume that the random vector is available and sampled at the end of production line or by a recursive online parameter estimator. The resulting partition {Λ q : q Z N } and its multiple controllers {K q : q Z N } are utilized by the selection function σ( ) : Λ Z N defined by the formula σ() = q, if Λ q for q Z N. (7) After finding the local region that contains an estimated parameter, a user can switch using the selection rule σ( ) in Eq. (7) from a robust controller for the entire uncertain region to a controller for the local region to improve performance. A. Partitioning the Polytopic Uncertainty Recall the uncertain parameter ( ) : Z + Λ. Suppose that Λ consists of two parameter subspaces, one of which is to be partitioned while the other is not, i.e., := T T T Λ R u, where Λ Rū, Λ R u and ū+u = u. The partitioning of Λ is assumed to be of importance to enhance the performance of the osedloop system, i.e., there is significant improvement if we partition Λ into multiple subsets. On the other hand, the parameter Λ is not measured or estimated, since it may not be possible to estimate this parameter and/or it may not significantly affect the robust performance of the osed-loop system. Now we present a useful technique for partitioning the entire uncertainty into a collection of convex polytopes: the Cartesian product partition technique. Cartesian Product Partition Technique Suppose the parameter set is contained in a hyperrectangular Λ ū min i i=, max i, (8) where Λ Rū. Let each osed interval be partitioned as min i = δ i < δi < < δi N i < δi N i = max i, (9) where N i is the number of sections in i th coordinate of Λ. Define the coordinate of the subsets in the partition (and its respective controllers) as a ū-tuple of positive integers where Z N is denoted by q := (q,, qū) Z N Zū+, (2) Z N := ū Z Ni, (2) i= to be the set of all ū-tuples of elements of q such that q i Z Ni := {, 2,, N i } for each i. The number of total subsets in the parametric uncertainty partition is N := N N 2 Nū. Finally, we define the intersection of the Cartesian product of the indexed family and Λ, denoted by Λ q ū Λ q := δq i, i δi q i Λ (22) i= to be the set of all ū-tuples of elements of Λ such that i δq i i, δq i i and i i, where i is the i th element of, for each i. Notice that Λ q := Λ q Λ is a nonempty osed convex set. Λ q s are illustrated in Fig. 2. By choosing appropriate δj i in Eq. (9), Λ q satisfies the conditions { in Eq. } (3). { The } elements in a subset of borders δ i k, : k Z Ni \ δ i, δn i i for i {,, ū} in Eq. (9) are the optimization parameters. B. Robust H Controller Design for a Fixed Partition For a given convex polytopic subset Λ q, a sub-optimal robust controller shall be designed. Let us parameterize the q th controller K q in Eq. () by defining a matrix Θ q as AKq B Θ q := Kq R C Kq D (n+nu) (n+ny). (23) Kq Then, as in the osed-loop system matrix with Λ q can be written in terms of Θ q A (Θ q ) B (Θ q) A := B C (Θ q) D (Θ q) C D B + Θ q C D2, D 2 (24)

δ δ δ 2 2 Λ Λ (,) Λ (,2) Λ(,3) Λ (2,) Λ (2,2) Λ (2,3) δ 2 δ 2 δ 2 2 δ 2 3 Fig. 2. A Cartesian product partition (solid line) and a locally updated partition (dotted line) for optimization. where A A := n n ; B n n := B ; n n n nw C := C nz n ; B n n B := 2 ; I n n n nu n n I C := n n ; D C 2 2 := nz n D2 ; ny n n nw D 2 :=. D 2 (25) Now we present an algorithm for finding a robustly stabilizing controller matrix Θ q that solves min max T zw (, Θ q ). (26) Θ q Λ q This can be solved by optimization over a finite number of matrix inequalities : min γ 2 {P q, : V(Λ q)},θ q,g q,γq 2 q, subject to M(P q,, Θ q, G q,, γ q ), for V(Λ q ), M(P q,, Θ q, G q,, γ q ) is given by (27) M(P q,, Θ q, G q,, γ q ) := P q, A (Θ q)g q B (Θ q) G q + Gq T P q, Gq TC (Θ q) T I D (Θ q) T, γqi 2 where represents entries which follow from symmetry and P q, is a matrix of appropriate size satisfying P q, = P T q,. G q is a matrix of the same size as P q,. A (Θ q), B (Θ q), C (Θ q) and D (Θ q) are defined in Eq. (24). Due to the uncertain parameter in the osedloop system matrix in Eq. (24) and Lemma, the osed-loop matrix as in Eq. (24) is convex for Λ q, therefore, the infinite number of matrix inequalities M(P q,, Θ q, G q,, γ q ), where Λ q become a finite number of matrix inequalities M(P q,, Θ q, G q,, γ q ), where V(Λ q ). In this way, we obtain the affine parameter dependent Lyapunov matrix defined as the convex combination of the uncertain parameters P q () := ξ i P q,i, (28) i V(Λ q) where ξ i are given by the following formula: Φ Λq := G : G = ξ ig i, ξ i, i V(Λ q) i V(Λ q) ξ i =, (29) where G and Φ Λq are obtained from Eqs. (9) and () respectively, when the system matrices in Eq. (7) are replaced by the osed loop system matrices in Eq. (24) for Λ q. This problem is however nonconvex, due to the coupling term between G q and Θ q in Eq. (27). A similar descent algorithm to the one used in 2, which utilizes the coordinate descent method for designing a H 2 controller, is adopted for finding sub-optimal H controllers. The procedure is as follows. ) Initial Design of Θ q : This is explained in Appendix. Set Θ to the result of the initial design. Also set j =. 2) Design of G q : Fix Θ q := Θ j q. Solve the convex optimization problem in Eq. (27) with respect to γ q, {P q,i : i V(Λ q )} and G q. Set Gq j to a solution G q. 3) Design of Θ q : Fix G q := Gq j. Solve the convex optimization problem in Eq. (27) with respect to γ q, {P q,i : i V(Λ q )} and Θ q. Set Θ j+ q to a solution Θ q. Increment j by one. Continue this iteration until γ q converges. Since the value γ q has a lower bound which is and is monotonically non-increasing during the iterations, it will converge to some positive number. A robust H controller covering the entire uncertainty Λ can be obtained in the same way with Eq. M(P, Θ, G,, γ), where V(Λ). C. Synthesis of a Partition with its Respective Set of Robust Controllers Our main problem is to solve the following optimization: ( { Λ opt q }, { Θ opt q } ) := arg min {Λ q},{θ q K(Λ q)} max J (Λ q,θ q). (3)

The following assumptions are introduced to develop an algorithm for solving the optimization problem in Eq. (3). B. A partition of the uncertainty set {Λ q } is wellchosen so that the optimization in Eq. (27) is feasible for a given subset Λ q. Moreover, the nonconvex optimization problem in Eq. (27) can be solved, and we denote ˆΘ q as the optimal solution for a given subset Λ q. B.2 We denote ˆγ q as the performance cost achieved in Eq. (27) by the optimal controller ˆΘ q : ˆγ q := max Λ q T zw (, ˆΘ q ). (3) Then, we assume that ˆγ q continuously changes with respect to the changes in the subset Λ q. The assumptions allow us to devise a descent synthesis algorithm to design a partition and multiple controllers. The feasibility of the assumptions will be discussed shortly. We now denote Λ j q as a given q th region of the partition at the iteration time j, and ˆΘ j q and ˆγj q (or ˆγ(Λj q )) as the corresponding optimal controller and resulting performance cost respectively. Then we have a following fact. Lemma 4: Let ˆγ : Λ R + be a function from a given partitioned subset in Λ to the optimized performance cost in R +, then the function ˆγ is continuous and monotone, i.e., whenever Λ j q Λ j+ q then ˆγ(Λ j q) ˆγ(Λ j+ q ). Proof: Continuity is given by assumption B.2. Monotonicity is guaranteed by the nature of the optimization, due to the inusion relationship Λ j q Λ j+ q. Now we present an algorithm to solve the main optimization problem in Eq. (3). ) Initial Design of {Λ q } and {Θ q }: Set { { } { } Λq} and Θ q to the initial designs. Also set γ q to the performance cost in Eq. (27) and set j =. 2) Updating { } {Λ q } and {Θ q } : Update the partition Λ j q so that the worst performance region should be reduced. Denote the set of coordinates of partitioned regions that produce the maximum total performance cost by { } L j := arg max γq j. (32) Take out as many elements as possible from these subsets { Λ j q : q L j} and assign them to the neighboring subsets by adjusting borders, subject to the constraint that any updated partitioned subset must be either Λ new q Λ j q, or Λ new q Λ j q, q Z N. (33) Do this until L new L j after redesigning regional robust controllers. Set { } { } Λ j+ q and Θ j+ q to the new partition and its controllers. Increment j by one. Continue this iteration until γ j q L j converges. In each iteration, we compute the set L j and update the partition and its controllers. In this way we can improve the total performance cost at each iteration. Remark 5: If the worst-case performance in the optimization problem in Eq. (3) is replaced by the averaged stochastic worst-case performance cost, min {Λ q},{θ q K(Λ q)} = min {Λ q},{θ q} min {Λ q},{θ q} = min {Λ q},{θ q} max EJ (Λ q, Θ q ) max max Λ q sup T zw (Σ, Θ q ) f ()d Λ q Λ q ˆγ q f ()d max ˆγ q Prob( Λ q ) }{{} :=J (Λ q,θ q) (34) the upper bound of this performance cost can be minimized via a similar algorithm when the probability density function f is continuous. J (Λ q, Θ q ) defined in Eq. (34) plays the same role of J (Λ q, Θ q ) in Eq. (3). Unfortunately, these optimization problems are nonconvex and assumptions B. may not be satisfied. Given the current state of robust control theory, we can not compute the optimal solution ˆΘ q for a given polytopic uncertain set Λ q 5. Therefore, in this paper, only a sub-optimal controller is computed as explained in the previous section. However, if the sub-optimal controller continuously changes with respect to the changes in the subset Λ q, then assumption B.2 holds, since the worst-case performance is the continuous function of the controller and the uncertainty simultaneously. Given the Cartesian product partition technique previously described, it is easy to update the partition subject to the inusion and exusion constraint Eq. (33) as shown in Fig. 2. For example, assume that q = (, 2) has the maximum performance cost among q Z N then in the next iteration, γ q=(,2) and γ q=(2,2) are expect to decrease while γ q=(,3) and γ q=(2,3) are expected to increase. The presented optimization problem in Eq. (6) is somewhat similar to the encoder and decoder design problem in the vector quantization theory 3. The encoder design is replaced by partitioning the compact parametrically uncertain set, and the decoder design is replaced by the optimization of a robust controller for each local subset. Finally, the usual quantization error cost is exchanged by the worst-case H 2 and/or H performance cost of the osed-loop system with respect to any parameter variations within the uncertainty.

2 Plot of regional robust controllers - -.3 7.958 (,) (,2) red-solid Magnitude (db) 6 5 4 (2,) (2,2) bluedashed greendashdot blackdotted Fig. 3. The final partition, its coordinates and a legend for the randomly generated parameters. Phase (deg) 3 27 225 8 V. AN ILLUSTRATIVE EXAMPLE Consider a second order mass-spring-damper system that often arises in micro-machined inertial sensor systems, as an illustrative example. The state space model is given by where A := ẋ(t) = Ax(t) + Bu(t) y(t) = Cx(t) + Dw(t), k m b m ; B := ; C := ; D := ; k m { 4 5 ( + w k ) : w k =.5, } b m {4( + w b 2) : w b =.5, 2 }. (35) (36) w k and w b represent 5% parameter variations. We select a performance weighting function W p (s) = g/(s + w c ), where w c = 9 2π and g = 8.4823 3 with the state space model ż(t) = w c z(t) + gy(t) to achieve a disturbance rejection : 5 for constant disturbances from w to y and a bandwidth of around 4 (rad/s) for all possible values of the uncertain parameter. This performance requirement can be also viewed as a good tracking response of the osed-loop system from the input signal (or the force on mass) to the output signal (or the position of mass). Combine Eq. (35) and W p (s) into a generalized plant as ẋ(t) ż(t) A 2 x(t) = gc w c z(t) B + u(t), z(t) = 2 x(t), z(t) y(t) = C 2 x(t) z(t) 2 + gd + w(t). w(t) (37) In order to test the algorithm for a discrete time system, the continuous time generalized plant in Eq. (37) 35 2 3 4 Frequency (rad/sec) Fig. 4. The single controller K single (black-solid line) and the multiple controllers {K q : q Z 4 } (see the legend in Fig. 3). Magnitude (db) Phase (deg) 5 5 8 35 9 45 45 9 35 8 Bode plot of perturbed open loop system u > y 225 2 3 Frequency (rad/sec) Fig. 5. Bode plot of the perturbed open-loop systems from u to y (see the legend in Fig. 3). with uncertain parameter channel, shown in Fig., was discretized with a sampling time Ts := 4. However the parameter uncertain set in the discertized system was kept the same as the one for the continuous time system as a matter of convenience 2. For this example, we found a single robust controller K single based on a parameter dependent Lyapunov function, achieving an upper bound of γ single = 2.378 for the entire uncertainty region using the algorithm explained in the section IV-B. Next we found a Cartesian prod- 2 We did not pursue the corresponding exact uncertain set in the discretized system transformed from the one in the continuous time system. In general, the corresponding uncertain set in the discretized system is not convex polytopic, even though the one in the continous time system is so. But with the fast sampling time, they are similar approximately.

Plot of osed loop sensitivity S(z) TABLE I WORST-CASE COSTS ACHIEVED BY THE MULTIPLE ROBUST CONTROLLERS. γ (,) =.8757 γ (,2) =.8754 γ (2,) =.875 γ (2,2) =.8758 2 Magnitude (db) 3 4 5 6 2 3 4 Frequency (rad/sec) Fig. 6. Sensitivity functions of the perturbed osed-loop systems (see the legend in Fig. 3) with the single controller. Magnitude (db) 2 3 4 5 6 Plot of osed loop sensitivity S(z) 2 3 4 Frequency (rad/sec) Fig. 7. Sensitivity functions of the perturbed osed-loop systems (see the legend in Fig. 3) with the multiple controllers. uct partition (two by two) as shown in Fig. 3 and its multiple controllers {K q : q Z N } using the optimization algorithm elaborated in the section IV-C. The previously found single controller K single served as initial controllers and the equally spaced partition was chosen as the initial partition for the nonconvex optimization as in Eq. (3). The optimization highly depends on initial conditions and the problem structure. The upper bounds achieved by multiple controllers are shown in Table I with respect to the coordinates of the partition. The iterations stopped when there were two big diagonal cost values such as γ (,) =.8757 and γ (2,2) =.8758, which blocks future improvement of the total cost by changing borders in any direction within specified tolerance. As shown in Table I, the total performance cost is γ total := max q Z4 γ q =.8758, which improves the cost obtained by the single controller γ single = 2.378 by 62%. The bode-diagram of the single controller and those of the four multiple controllers are plotted in Fig. 4. It is very interesting to notice that K (,) K (,2) and K (2,) K (2,2), which hints the enhanced significance of partitioning the space (or spring constant parameter space) as compared to the 2 space (or damping coefficient parameter space). In order to evaluate the osed-loop uncertain systems, twenty perturbed systems were generated by parameters sampled from a stochastic process with a probability density function f (ν, ν 2 ) = f (ν )f 2 (ν 2 ), where f (ν) = f 2 (ν) = /2, if ν, otherwise f (ν) = f 2 (ν) =, (i.e., it consists of two independent uniform probability density functions). Each randomly generated open-loop plant from the input signal u to the output signal y is assified and plotted according to the subset of the partition where it is contained as shown in Fig. 5. The six plants that belong to Λ (,) are plotted with red solid lines, the five plants that belongs to Λ (,2) are plotted with blue dashed lines, the four plants that belong to Λ (2,) are plotted with green dash-dotted lines, and the five plants that belong to Λ (2,2) are potted with black dotted lines, as summarized in Fig. 3. These osed-looped sensitivity functions S(z) from the disturbance d to the controlled output y by the single controller K single are shown in Fig. 6 as compared to the corresponding multiple controllers {K q : q Z 4 } which are shown in Fig. 7. It is ear to see that the sampled sensitivity function transfer functions by the partition and its multiple controllers, as shown in Fig. 7, are better shaped as compared to the ones with a single controller as shown in Fig. 6. VI. CONCLUSIONS In this paper, we presented an algorithm to simultaneously determine a convex polytopic partitioning of an uncertainty region, and to design a set of robust controllers for LTI systems with parametric uncertainty Λ, where Λ is convex polytope. Each controller takes charge of a local region and is designed, based on a parameter dependent Lyapunov function, to give a sub-optimal worst-cast performance for that region. Minimization of the maximum of such worst-case performance of each local region is considered with respect to a partition and its multiple robust controllers. The presented algorithm updates both partitions and controllers to achieve monotonic non-increasing

total performance. An illustrative example was studied to evaluate the algorithm and to demonstrate the benefits of partitioning the uncertainty space and utilizing multiple robust controllers. As for future work, online/offline parameter estimation algorithms for the correct selection shall be explored as well as improved partitioning algorithms. Extensions from LTI parametric uncertain systems to linear parameter-varying (LPV) systems are quite natural and currently under investigation. APPENDIX INITIAL CONTROLLER DESIGN IN ROBUST H SYNTHESIS For nonconvex optimization, the choice of an initial point is critical. We review a reasonable method for selection of an initial point proposed in 5. A. State Feedback First, for the uncertain system Eq. (5), we design a state feedback controller: u = Fx that optimizes H norm of the osed loop system: sup Λq T zw. As in Theorem 6 in 5, this can be solved by convex optimization with LMIs. min γ 2 Q,L,γq 2 q subject to Q A B Q C T I D T γqi 2, (38) where A := A Q + B 2 L and C := C Q + D 2L. The constraints are imposed at the vertices of Λ q. The optimal state feedback is given by F := LQ, (39) where L R nu n and Q = Q T R n n are the solutions of the optimization problem in Eq. (38). B. Output Feedback Fixing F in Eq. (39) as C Kq and D Kq = in Eq. (23), we design a full order dynamic output feedback AKq B u = Kq y, where A F Kq R n n (4) that optimizes H performance. Due to Theorem 7 in 5, this problem can be solved by convex optimization with LMIs: min γ 2 X,Y,Z,G,γq 2 q, subject to P PÃ P B T P C T I D, γqi 2 (4) where P := X Y R 2n 2n, X = X T R n n, Y = Y T R n n, Z := Y A K R n n and G := Y B K R n ny. Moreover, the matrices PÃ, P B, C and D are defined as follows PÃ := X(A + B2 F) XB2 F Y (A + B2 F) Z GC 2 Z Y B2 F, P B XB := Y B GD2, D := D, C := C + D2F D2F. Again, the constraints in the LMIs in Eq. (4) are imposed at the vertices of Λ q. Using the optimizers, A Kq and B Kq are obtained as A Kq := Y Z, B Kq := Y G. (42) REFERENCES Geir E. Dullerud and Fernando Paganini. A Course in Robust Control Theory: A Convex Approach, volume 36 of Texts in Applied Mathematics. Springer-Verlag, 2. 2 B. D. O. Anderson, T. S. Brinsmead, F. de Bruyne, J. P. Hespanha, D. Liberzon, and A. S. Morse. Multiple model adaptive control. I. finite controller coverings. International Journal of Robust and Nonlinear Control, (-2):99 929, September 2. 3 J. P. Hespanha, D. Liberzon, A. S. Morse, B. D. O. Anderson, T. S. Brinsmead, and F.de Bruyne. Multiple model adaptive control, part 2: Switching. International journal of robust and nonlinear control, (5):479, 2. 4 A. Stephen Morse. Control using logic-based switching. Lecture notes in control and information sciences ; 222. Springer, London ; New York, 997. 5 S. Kanev, C. Scherer, M. Verhaegen, and B. De Shutter. Robust output-feedback controller design via local BMI optimization. Automatica, 4:5 27, 24. 6 Lennard Ljung. Theory and Practice of Recursive Identification. The MIT Press, Cambridge, Massachusetts, 983. 7 Greg Wolodkin, Sundeep Rangan, and Kameshwar Poollar. An LFT approach to parameter estimation. In Proceedings of the American Control Conference, pages 288 2284, 997. 8 Pascal Gahinet, Arkadi Nemirovski, Alan J. Laub, and Mahmoud Chilali. LMI Control Toolbox User s Guide. The Mathworks, 995. 9 Carsten Scherer, Pascal Gahinet, and Mahmoud Chilali. Multiobjective output-feedback control via LMI optimization. IEEE Transactions on Automatic Control, 42:896 9, 997. Pascal Gahinet and Pierre Apkarian. A linear matrix inequality approach to H control. International Journal of Robust and Nonlinear Control, 4:42 448, 994. M. C. De Oliveira, J. C. Geromel, and J. Bernussou. Extended H 2 and H norm characterizations and controller parametrizations for discrete-time systems. International Journal of Control, 75(9):666 679, June 22. 2 Ryozo Nagamune, Xinghui Huang, and Roberto Horowitz. Robust control synthesis techniques for multirate and multi-sensing trackfollowing servo systems in HDDs. (In preparation), 25. 3 Allen Gersho and Robert M. Gray. Vector Quantization and Signal Compression. Kluwer Academic Publishers, 2.